[6/7] vdpa/mlx5: Avoid using reslock in event_handler

  event_handler runs under atomic context and may not acquire reslock. We
can still guarantee that the handler won't be called after suspend by
clearing nb_registered, unregistering the handler and flushing the
workqueue.

Signed-off-by: Eli Cohen <elic@nvidia.com>
---
 drivers/vdpa/mlx5/net/mlx5_vnet.c | 14 +++-----------
 1 file changed, 3 insertions(+), 11 deletions(-)
  

On Sun, Nov 13, 2022 at 9:45 PM Eli Cohen <elic@nvidia.com> wrote:
>
> event_handler runs under atomic context and may not acquire reslock. We
> can still guarantee that the handler won't be called after suspend by
> clearing nb_registered, unregistering the handler and flushing the
> workqueue.
>
> Signed-off-by: Eli Cohen <elic@nvidia.com>
> ---
>  drivers/vdpa/mlx5/net/mlx5_vnet.c | 14 +++-----------
>  1 file changed, 3 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> index 6e6490c85be2..bebfba530247 100644
> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> @@ -2872,8 +2872,8 @@ static int mlx5_vdpa_suspend(struct vdpa_device *vdev)
>         int i;
>
>         down_write(&ndev->reslock);
> -       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
>         ndev->nb_registered = false;
> +       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);

I wonder why this can help anything. And if it does, we have simliar
logic in mlx5_vdpa_dev_del() do we need to fix that as well?

Thanks

>         flush_workqueue(ndev->mvdev.wq);
>         for (i = 0; i < ndev->cur_num_vqs; i++) {
>                 mvq = &ndev->vqs[i];
> @@ -3051,7 +3051,7 @@ static void update_carrier(struct work_struct *work)
>         else
>                 ndev->config.status &= cpu_to_mlx5vdpa16(mvdev, ~VIRTIO_NET_S_LINK_UP);
>
> -       if (ndev->config_cb.callback)
> +       if (ndev->nb_registered && ndev->config_cb.callback)
>                 ndev->config_cb.callback(ndev->config_cb.private);
>
>         kfree(wqent);
> @@ -3068,21 +3068,13 @@ static int event_handler(struct notifier_block *nb, unsigned long event, void *p
>                 switch (eqe->sub_type) {
>                 case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
>                 case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
> -                       down_read(&ndev->reslock);
> -                       if (!ndev->nb_registered) {
> -                               up_read(&ndev->reslock);
> -                               return NOTIFY_DONE;
> -                       }
>                         wqent = kzalloc(sizeof(*wqent), GFP_ATOMIC);
> -                       if (!wqent) {
> -                               up_read(&ndev->reslock);
> +                       if (!wqent)
>                                 return NOTIFY_DONE;
> -                       }
>
>                         wqent->mvdev = &ndev->mvdev;
>                         INIT_WORK(&wqent->work, update_carrier);
>                         queue_work(ndev->mvdev.wq, &wqent->work);
> -                       up_read(&ndev->reslock);
>                         ret = NOTIFY_OK;
>                         break;
>                 default:
> --
> 2.38.1
>
  

> From: Jason Wang <jasowang@redhat.com>
> Sent: Monday, 14 November 2022 9:53
> To: Eli Cohen <elic@nvidia.com>
> Cc: mst@redhat.com; linux-kernel@vger.kernel.org; virtualization@lists.linux-
> foundation.org; si-wei.liu@oracle.com; eperezma@redhat.com;
> lulu@redhat.com
> Subject: Re: [PATCH 6/7] vdpa/mlx5: Avoid using reslock in event_handler
> 
> On Sun, Nov 13, 2022 at 9:45 PM Eli Cohen <elic@nvidia.com> wrote:
> >
> > event_handler runs under atomic context and may not acquire reslock. We
> > can still guarantee that the handler won't be called after suspend by
> > clearing nb_registered, unregistering the handler and flushing the
> > workqueue.
> >
> > Signed-off-by: Eli Cohen <elic@nvidia.com>
> > ---
> >  drivers/vdpa/mlx5/net/mlx5_vnet.c | 14 +++-----------
> >  1 file changed, 3 insertions(+), 11 deletions(-)
> >
> > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > index 6e6490c85be2..bebfba530247 100644
> > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > @@ -2872,8 +2872,8 @@ static int mlx5_vdpa_suspend(struct vdpa_device
> *vdev)
> >         int i;
> >
> >         down_write(&ndev->reslock);
> > -       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
> >         ndev->nb_registered = false;
> > +       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
> 
> I wonder why this can help anything.
I think you were concerned that async events will come when the device was suspended. Since we can't take reslock, I think this guarantees that we won't get any events after suspension.

> And if it does, we have simliar
> logic in mlx5_vdpa_dev_del() do we need to fix that as well?
> 
We have the same construct there only that I set nb_registered = false after unregistering the notifier. So I probably need to move it before mlx5_notifier_unregister().

> Thanks
> 
> >         flush_workqueue(ndev->mvdev.wq);
> >         for (i = 0; i < ndev->cur_num_vqs; i++) {
> >                 mvq = &ndev->vqs[i];
> > @@ -3051,7 +3051,7 @@ static void update_carrier(struct work_struct
> *work)
> >         else
> >                 ndev->config.status &= cpu_to_mlx5vdpa16(mvdev,
> ~VIRTIO_NET_S_LINK_UP);
> >
> > -       if (ndev->config_cb.callback)
> > +       if (ndev->nb_registered && ndev->config_cb.callback)
> >                 ndev->config_cb.callback(ndev->config_cb.private);
> >
> >         kfree(wqent);
> > @@ -3068,21 +3068,13 @@ static int event_handler(struct notifier_block
> *nb, unsigned long event, void *p
> >                 switch (eqe->sub_type) {
> >                 case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
> >                 case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
> > -                       down_read(&ndev->reslock);
> > -                       if (!ndev->nb_registered) {
> > -                               up_read(&ndev->reslock);
> > -                               return NOTIFY_DONE;
> > -                       }
> >                         wqent = kzalloc(sizeof(*wqent), GFP_ATOMIC);
> > -                       if (!wqent) {
> > -                               up_read(&ndev->reslock);
> > +                       if (!wqent)
> >                                 return NOTIFY_DONE;
> > -                       }
> >
> >                         wqent->mvdev = &ndev->mvdev;
> >                         INIT_WORK(&wqent->work, update_carrier);
> >                         queue_work(ndev->mvdev.wq, &wqent->work);
> > -                       up_read(&ndev->reslock);
> >                         ret = NOTIFY_OK;
> >                         break;
> >                 default:
> > --
> > 2.38.1
> >
  

On Mon, Nov 14, 2022 at 4:58 PM Eli Cohen <elic@nvidia.com> wrote:
>
> > From: Jason Wang <jasowang@redhat.com>
> > Sent: Monday, 14 November 2022 9:53
> > To: Eli Cohen <elic@nvidia.com>
> > Cc: mst@redhat.com; linux-kernel@vger.kernel.org; virtualization@lists.linux-
> > foundation.org; si-wei.liu@oracle.com; eperezma@redhat.com;
> > lulu@redhat.com
> > Subject: Re: [PATCH 6/7] vdpa/mlx5: Avoid using reslock in event_handler
> >
> > On Sun, Nov 13, 2022 at 9:45 PM Eli Cohen <elic@nvidia.com> wrote:
> > >
> > > event_handler runs under atomic context and may not acquire reslock. We
> > > can still guarantee that the handler won't be called after suspend by
> > > clearing nb_registered, unregistering the handler and flushing the
> > > workqueue.
> > >
> > > Signed-off-by: Eli Cohen <elic@nvidia.com>
> > > ---
> > >  drivers/vdpa/mlx5/net/mlx5_vnet.c | 14 +++-----------
> > >  1 file changed, 3 insertions(+), 11 deletions(-)
> > >
> > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > index 6e6490c85be2..bebfba530247 100644
> > > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > @@ -2872,8 +2872,8 @@ static int mlx5_vdpa_suspend(struct vdpa_device
> > *vdev)
> > >         int i;
> > >
> > >         down_write(&ndev->reslock);
> > > -       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
> > >         ndev->nb_registered = false;
> > > +       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
> >
> > I wonder why this can help anything.
> I think you were concerned that async events will come when the device was suspended. Since we can't take reslock, I think this guarantees that we won't get any events after suspension.
>

Ok, I see.

> > And if it does, we have simliar
> > logic in mlx5_vdpa_dev_del() do we need to fix that as well?
> >
> We have the same construct there only that I set nb_registered = false after unregistering the notifier. So I probably need to move it before mlx5_notifier_unregister().

Right.

Thanks

>
> > Thanks
> >
> > >         flush_workqueue(ndev->mvdev.wq);
> > >         for (i = 0; i < ndev->cur_num_vqs; i++) {
> > >                 mvq = &ndev->vqs[i];
> > > @@ -3051,7 +3051,7 @@ static void update_carrier(struct work_struct
> > *work)
> > >         else
> > >                 ndev->config.status &= cpu_to_mlx5vdpa16(mvdev,
> > ~VIRTIO_NET_S_LINK_UP);
> > >
> > > -       if (ndev->config_cb.callback)
> > > +       if (ndev->nb_registered && ndev->config_cb.callback)
> > >                 ndev->config_cb.callback(ndev->config_cb.private);
> > >
> > >         kfree(wqent);
> > > @@ -3068,21 +3068,13 @@ static int event_handler(struct notifier_block
> > *nb, unsigned long event, void *p
> > >                 switch (eqe->sub_type) {
> > >                 case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
> > >                 case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
> > > -                       down_read(&ndev->reslock);
> > > -                       if (!ndev->nb_registered) {
> > > -                               up_read(&ndev->reslock);
> > > -                               return NOTIFY_DONE;
> > > -                       }
> > >                         wqent = kzalloc(sizeof(*wqent), GFP_ATOMIC);
> > > -                       if (!wqent) {
> > > -                               up_read(&ndev->reslock);
> > > +                       if (!wqent)
> > >                                 return NOTIFY_DONE;
> > > -                       }
> > >
> > >                         wqent->mvdev = &ndev->mvdev;
> > >                         INIT_WORK(&wqent->work, update_carrier);
> > >                         queue_work(ndev->mvdev.wq, &wqent->work);
> > > -                       up_read(&ndev->reslock);
> > >                         ret = NOTIFY_OK;
> > >                         break;
> > >                 default:
> > > --
> > > 2.38.1
> > >
>
  

On Wed, Nov 23, 2022 at 12:00 PM Jason Wang <jasowang@redhat.com> wrote:
>
> On Mon, Nov 14, 2022 at 4:58 PM Eli Cohen <elic@nvidia.com> wrote:
> >
> > > From: Jason Wang <jasowang@redhat.com>
> > > Sent: Monday, 14 November 2022 9:53
> > > To: Eli Cohen <elic@nvidia.com>
> > > Cc: mst@redhat.com; linux-kernel@vger.kernel.org; virtualization@lists.linux-
> > > foundation.org; si-wei.liu@oracle.com; eperezma@redhat.com;
> > > lulu@redhat.com
> > > Subject: Re: [PATCH 6/7] vdpa/mlx5: Avoid using reslock in event_handler
> > >
> > > On Sun, Nov 13, 2022 at 9:45 PM Eli Cohen <elic@nvidia.com> wrote:
> > > >
> > > > event_handler runs under atomic context and may not acquire reslock. We
> > > > can still guarantee that the handler won't be called after suspend by
> > > > clearing nb_registered, unregistering the handler and flushing the
> > > > workqueue.
> > > >
> > > > Signed-off-by: Eli Cohen <elic@nvidia.com>
> > > > ---
> > > >  drivers/vdpa/mlx5/net/mlx5_vnet.c | 14 +++-----------
> > > >  1 file changed, 3 insertions(+), 11 deletions(-)
> > > >
> > > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > > index 6e6490c85be2..bebfba530247 100644
> > > > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > > @@ -2872,8 +2872,8 @@ static int mlx5_vdpa_suspend(struct vdpa_device
> > > *vdev)
> > > >         int i;
> > > >
> > > >         down_write(&ndev->reslock);
> > > > -       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
> > > >         ndev->nb_registered = false;
> > > > +       mlx5_notifier_unregister(mvdev->mdev, &ndev->nb);
> > >
> > > I wonder why this can help anything.
> > I think you were concerned that async events will come when the device was suspended. Since we can't take reslock, I think this guarantees that we won't get any events after suspension.
> >
>
> Ok, I see.
>
> > > And if it does, we have simliar
> > > logic in mlx5_vdpa_dev_del() do we need to fix that as well?
> > >
> > We have the same construct there only that I set nb_registered = false after unregistering the notifier. So I probably need to move it before mlx5_notifier_unregister().
>
> Right.
>
> Thanks

So I'm fine with this patch.

Acked-by: Jason Wang <jasowang@redhat.com>

Thanks

>
> >
> > > Thanks
> > >
> > > >         flush_workqueue(ndev->mvdev.wq);
> > > >         for (i = 0; i < ndev->cur_num_vqs; i++) {
> > > >                 mvq = &ndev->vqs[i];
> > > > @@ -3051,7 +3051,7 @@ static void update_carrier(struct work_struct
> > > *work)
> > > >         else
> > > >                 ndev->config.status &= cpu_to_mlx5vdpa16(mvdev,
> > > ~VIRTIO_NET_S_LINK_UP);
> > > >
> > > > -       if (ndev->config_cb.callback)
> > > > +       if (ndev->nb_registered && ndev->config_cb.callback)
> > > >                 ndev->config_cb.callback(ndev->config_cb.private);
> > > >
> > > >         kfree(wqent);
> > > > @@ -3068,21 +3068,13 @@ static int event_handler(struct notifier_block
> > > *nb, unsigned long event, void *p
> > > >                 switch (eqe->sub_type) {
> > > >                 case MLX5_PORT_CHANGE_SUBTYPE_DOWN:
> > > >                 case MLX5_PORT_CHANGE_SUBTYPE_ACTIVE:
> > > > -                       down_read(&ndev->reslock);
> > > > -                       if (!ndev->nb_registered) {
> > > > -                               up_read(&ndev->reslock);
> > > > -                               return NOTIFY_DONE;
> > > > -                       }
> > > >                         wqent = kzalloc(sizeof(*wqent), GFP_ATOMIC);
> > > > -                       if (!wqent) {
> > > > -                               up_read(&ndev->reslock);
> > > > +                       if (!wqent)
> > > >                                 return NOTIFY_DONE;
> > > > -                       }
> > > >
> > > >                         wqent->mvdev = &ndev->mvdev;
> > > >                         INIT_WORK(&wqent->work, update_carrier);
> > > >                         queue_work(ndev->mvdev.wq, &wqent->work);
> > > > -                       up_read(&ndev->reslock);
> > > >                         ret = NOTIFY_OK;
> > > >                         break;
> > > >                 default:
> > > > --
> > > > 2.38.1
> > > >
> >